Dynamic Compact Thermal Model with Neural Networks for Radar Applications

HAL Id: hal-00171366

https://hal.archives-ouvertes.fr/hal-00171366

Submitted on 12 Sep 2007

HAL is a multi-disciplinary open access archive for the deposit and dissemination of sci- entific research documents, whether they are pub- lished or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers.

L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires publics ou privés.

Dynamic Compact Thermal Model with Neural Networks for Radar Applications

G. Mallet, Ph. Leray, H. Polaert, C. Tolant, Ph. Eudeline

To cite this version:

G. Mallet, Ph. Leray, H. Polaert, C. Tolant, Ph. Eudeline. Dynamic Compact Thermal Model with Neural Networks for Radar Applications. THERMINIC 2006, Sep 2006, Nice, France. pp.118-122.

�hal-00171366�

Dynamic Compact Thermal Model with Neural Networks for Radar Applications

Gr´egory Mallet

, Philippe Leray

, Hubert Polaert

, Cl´ement Tolant

, Philippe Eudeline

1

LITIS EA 4051 - INSA de Rouen

2

Advanced studies - THALES AIR DEFENCE Rouen

ABSTRACT

This article deals with the creation of a compact thermal model. In this aim, we apply some well-known methods such as FEM model reduction and identification of RC networks. To go further than already existing approaches, we also introduce the use of artificial neural networks (ANNs) to cope with nonlinearities which may appear in thermal phenomenons. A new hybrid model, trying to gather the advantages of ANNs and RC networks, is ap- plied on a simple thermal problem. The need of samples will also lead us to carry out, in parallel, the FEM model reduction. The reduced FEM model will then be used to generate the required databases and validate the compact model results.

1. INTRODUCTION

The increase of power density and the constant miniatur- ization of electronic elements have generated an growing requirement of thermal management. Thus, in active an- tenna, the junction temperature of power modules, typ- ically high power amplifiers (HPA), should be carefully evaluated. Indeed, this temperature affects both the reli- ability and the efficiency of the components and taking into account this parameter becomes more and more im- portant. The Finite Elements Method (FEM) seems to bring a solution for this problem. Indeed, a fine mod- eling makes possible to predict quite accurately the heat repartition inside all kinds of components and/or cool- ing systems. Nevertheless, build a fully-detailled model is impossible and the resolution of the numerical system can be really time-consuming, mainly if the model is ac- curate and hence complex.

Lots of tips are of course available to reduce the com- plexity of an FEM model. Using symmetries of the model or neglecting some thermal effects could simplify the problem and accelerate calculations. Those techniques could be very useful, but the FEM offently stays too slow to be really efficient. A step forward in the reduction of complexity is done with Compact Thermal Models (CTMs). Those methods have no direct link with the real structure of the system and only preserve a notion

of ”level”. A complete part of the system is then often replaced by a single node in an equivalent electronic cir- cuit. The parameter of the circuit are calculated to imitate the response of the real system. Many studies have been achieved on this way, and the results have shown CTMs could be a very attractive method for thermal prediction [1] [2] [3].

For dynamic responses, and when the conduction is not the main thermal phenomenon, the nonlinearities of the real systems are however more perceptible and some improvements of CTMs seem possible. Given the fact that this method can be viewed as a linear model identifi- cation, it appears sensible to use nonlinear models identi- fication (such as ANN) in order to surmount current lim- itations of CTMs.

To cover all the aspects of this question, we will re- duce the real case of power modules to a simpler one with only one electronic component mounted on a cooling sys- tem (Fig 1). The problem of temperature prediction is hence much easier but the thermal effects involved are similar. The FEM model of the system will be used to evaluate each steps of the compact model creation and to provide simulated measurements. As all the others in the article, this FEM model has been build with TAS (Ther- mal Analysis System) Software.

Figure 1: Global view of the system.

2. RC NETWORK AS COMPONENT MODEL

The FEM model is voluntarily simple. The junction tem- perature prediction is thus very coarse and low. The model is only composed of a silicon bar on a block of BeO brased on a support made of CuW (Fig 2). The sup- port will next be used to screwed the ”component” on the radiator.

Figure 2: FEM model of the component

For this model, the main thermal effect is due to con- duction. In this case, for small dimension elements, some assumptions could be very useful to simplify the model.

If the heat flow is supposed to be one-way, the thermal behaviour of a homogeneous block of material can be summarized by its thermal resistance R

and its thermal capacitance C

.

R

= ∆x

λS (1)

C

= µcS∆x (2)

∆x : length of the block in the direction of heat flow λ : conductivity of the material

S : surface perpendicular to the direction of heat flow µ : volumetric mass of the material

c : specific heat of the material

The dynamic response of the component can then be reproduced by a network with 3 RC elements (one for each homogenous block). To identify thermal resistances, the static response for a fixed boundary condition is sufficient (Fig 3). However, the calculation of capacitances requires the temperature transcients (Fig 4) at each level. Thanks to FEM, it is possible to easily extract all those data. The parameters of an equivalent model can so be computed. This kind of approach is used by the majority of the CTMs methods [4] [5] [6].

3. FEM MODEL REDUCTION

TAS allowes the introduction of a RC network inside the structure of a model. Hence, the equivalent circuit can be inserted in the FEM model. The substitution with the true

Figure 3: Temperature of the component in steady case

Figure 4: Temperature curves for a 20 W power-step

component model saves about 700 nodes. But the major- ity of nodes are still dedicated to model fins. A second simplification is also available on the FEM model. The convective effects are characterized by a heat exchange coefficient which is linked to surface area. Thus, a plate with fins is equivalent to a simple plate with the same global heat exchange coefficient. This modification is however not sufficient. Indeed, the fins also change the conductivity of the cooling system. So, the parameters of the equivalent plate can not stay isotropic. The use of this replacement also enables us to reduce the numbers of convective flows.

The numbers of nodes has thus decreased from about 16000 for the initial complete model to less than 1000 for the simpliest one, with only a small deformation on the heat repartition (Fig 5 and 6). The latter will help us to generate databases for the learning of the neural network.

4. ANN COMPACT MODEL

Even if RC networks have shown their efficiency, they are

not applicable in all cases. For the cooling system, the

heat direction could not be considered one-way. More-

over, at first, the conduction is the main thermal phe-

nomenon in the radiator but, after a while, the tempera-

Figure 5: Repartition on heat exchanger with real com- ponent model and fins

Figure 6: Repartition on heat exchanger with equivalent RC network and without fins

ture in the radiator increases and therefore the heat ex- change coefficient. The convection then becomes pre- dominant. So, the radiator acts initially like a simple plate and stores the heat. When it reaches a specific tempera- ture, the heat load can then be evacuated. The cooling system could thus not be replaced by a simple RC circuit between the bottom of the component and the ambient temperature (Fig 7).

ANNs are able to accomplish this task of nonlinear regression. They have indeed been applied on various problems of pattern recognition, temporal series predic- tion or modelisation [7] [8]. A neuron is composed of several entries and one ouput. The output of the neuron is calculated with an activation function (usually non lin- ear) from an aggregation of the inputs. For our model, a single layer perceptron, the aggregation function is a weighted-sum and the activation function is the logistic function. The parameters of the network (the weights of the sums) are at first randomly initialiazed, and an out- put is computed. Then, a gradient descent method, which is called the learning algorithm, calculates the parame- ters correction that leads to a lowest error. This step is repeated until a specific error.

Figure 7: FEM model and RC model responses compari- sion for the radiator

ANNs are often presented as black-box models.

However, if their parameters are not linked to the real system, physics thoughts have a prominent weight in the choice of the neural network structure. The character- istics of the modeling problems force us to choose a specific form of dynamic neural network, called NNOE (Neural Network Output Error) [9]. NNOE use multi- delayed signals and their past outputs as inputs. The value of the time-step in the network should be care- fully selected in function of system dynamics. The pre- viously calculated RC network is really helpful in this aim. The circuit is indeed a low-pass filter and gives us directly the maximum frequency of power pulses that can be transmitted to the cooling system. Endly, the di- mensions of the radiator and the conductivity of the alu- minium bounds the size of cooling system ”memory” and thus the number of delays. The ANN is simulated with MATLAB and the NNSYSID toolbox.

For our ANN model, the only input is the mean power

injected in the component (Fig 8). After the physical

analysis, some parameters still remain to be selected. The

number of delays have been bound but the neural network

should rarely need all information. The input delays have

then been limited to 10. For stability reasons, the number

of delays on the reinjected output has been fixed to 1. Ten

neurons are located in the hidden layer. This parameter

can be analysed as a measure of the nonlinearity degree

of the system, but it should also be related to the num-

ber of inputs. The reduced FEM model has been used to

generate 3 measurements of 1 minut (learning, validation

and test databases) of the temperature under the compo-

nent for random power signals (only the active periods

are the sames). The results are compared with the FEM

model (Fig 9).

Figure 8: The real power peaks used in the RC model and the mean power, the input for the ANN

Figure 9: Output comparison of neural network model and FEM model for the temperature under the component

5. HYBRID MODEL AND RESULT

The merging of the two compacts models is quite triv- ial. For the ANN, the RC network acts as a filter on the power pulses, so the variation of the junction tem- perature will not impact the output of the neural network.

For the RC network, the neural network output is just a boundary condition fixed under the component during a time-step. The two models can then be considered as in- dependant and the junction temperature will just be the sum of their outputs. The only drawback of this method is on the update of the neural network output where a kind of ”discontinuity” may occur on the predicted junc- tion temperature signal. The hybrid model is then just the superposition of the ouputs of the 2 compact models.

A new set of data is generated to compare the out- put of the compact model and of the FEM model for both high and low frequencies (Fig 10). As the two parts of the model predict accurately their component parts of the signal, the combination of their outputs obviously pre-

dicts quite well the junction temperature.

Figure 10: Output comparison of hybrid model and FEM model for the junction temperature

6. CONCLUSIONS

The ANN compact model has shown its efficiency. The neural networks have then been able to deal with thermal non-linearities, like the convection in fins. Added to the RC network effective method, ANN and other statistical models, which will be tried in further works, provide a nice opportunity to improve compact models. The first improvement will be to use this method in the multiple components case.

Moreover, the hybrid model has been learnt only from data, even if some parameters have to be manually se- lected. Since there is no link between the physical model and the compact model, the latter can also be learnt from measurements and this method could create directly a model from a real system. Furthermore, neural networks can be adapted for control applications. As linear models [10], the hybrid model can so be used to limit the junction temperature. Another advantage of neural networks is to allow multiple inputs problem and thus some other vari- ables, such as ambient temperature or air velocity, could be taken into account.

7. REFERENCES

[1] C.J.M. Lasance. Recent progress in compact ther- mal models. In SEMITHERM XIX, pages 290–299, 2003.

[2] J. Palacin, M. Salleras, J. Samitier, and S. Marco.

Dynamic compact thermal models with multiple

power sources: application to an ultrathin chip

stacking technology. Advanced Packaging, IEEE

transaction, 28(4):694–703, 2005.

[3] G. Farkas, A. Poppe, E. Kollar, and P. Stehouwer.

Dynamic compact models of cooling mounts for fast board level design. In SEMITHERM XIX, pages 255–262, 2003.

[4] W. Huang, M. R. Stan, and K. Skadron. Physically- based compact thermal modeling-achieving param- eterization and boundary condition independence.

In THERMINICS, pages 287–292, 2004.

[5] M.Rencz, V.Sz´ekely, A.Poppe, and B.Courtois. In- clusion of rc compact models of packages into board level thermal simulation tools. SEMI- THERM, 2002.

[6] Marta Rencz and Vladimir Sz´ekely. Non-linearity issues in the dynamic compact model generation. In SEMI-THERM XIX, 2003.

[7] Coryn A.L. Bailer-Jones, David J.C. MacKay, and Philip J. Withers. A recurrent neural network for modelling dynamical systems. Network: Computa- tion in Neural Systems, 9:531 – 547, 1998.

[8] M. Hallas and G. Dorffner. A comparative study on feedforward and recurrent neural networks in time series prediction using gradient descent learning. In R. Trappl, editor, Cybernetics and Systems 98, Pro- ceedings of 14th European Meeting on Cybernet- ics and Systems Research, Vienna, pages 644–647, 1998.

[9] M. Norgard, O. Ravn, NK Poulsen, and LK Hansen.

Neural Networks for Modelling and Control of Dy- namics Systems. Springer-Verlag, 2000.

[10] Kevin Skadron, Tarek Abdelzaher, and Mircea R.

Stan. Control-theoretic techniques and thermal-RC

modeling for accurate and localized dynamic ther-

mal management. Technical Report CS-2001-27,

2001.